Fig. 4

Visualization of DeepResolve in multi-task networks. a Overall Feature Importance Vector for Synthetic dataset II class 1 - 4. Each circle on the X-axis represents a channel, with red representing positive OFIV score and blue representing negative OFIV score. Each column corresponds to one of the 32 channels that is shared among all four classes. OFIV successfully ranks predefined sequence features as the most important features for each of the classes, while reveals ‘unfavored’ features that are used to separate a class from its competing classes. b Correlation matrix of class based features shows the benefit of non-negative OFIV scores. The predefined sequence pattern for each class is shown (a). Our proposed Class Similarity Matrix (top-left) successfully assigns high correlation to (Class1, Class2), (Class2, Class3) and (Class1, Class3) and low correlation to all pairs with Class 4. The matrix in top right corner suggest low correlation between the labels of each class. The matrix on the bottom left is the Pearson correlation of ONIV score without removing the negative terms, and the bottom right matrix is calculated by taking the cosine of the corresponding rows in last layer weight matrix. The bottom two both fail to assign higher similarity score to combinations of classes that share sequence features